EdgeAgentX-DT：将数字孪生与生成式AI结合用于战术网络中的弹性边缘智能

Dr. Abir Ray∗ ∗Cornell UniversityEmail: ar2486@cornell.edu and decentralized execution, these edge agentscoordinatetheir strategies and outperform independent learners. Built-in adversarial AI defenses (robust aggregation, adversarialtraining, secure protocols) further ensure that learning remainsstable under jamming or model poisoning attacks. Experimen-tal evaluations showed EdgeAgentX achieving significantlylower latency, higher throughput, and faster convergence thanbaseline approaches (independent RL, standard FL, etc.), whilemaintaining resilience with minimal performance degradationunder simulated attacks. Abstract—We present a full-fledged extension of theEdgeAgentX framework by incorporating digital twin simula-tionsand generative AI-driven scenario training to enhanceedge intelligence in military networks. EdgeAgentX-DT leveragesnetwork digital twins – virtual replicas of the tactical network– synchronized with real-world edge devices to provide a safe,high-fidelity environment for training and testing. On top of this,generative AI (GenAI) techniques (including diffusion modelsand transformers) generate diverse and adversarial scenarios forthe agents to experience in simulation. This combined approachenables the edge agents to learn robust communication and coor-dination strategies under a wide spectrum of conditions withoutrisking mission assets. We detail the multi-layer architecture ofEdgeAgentX-DT consisting of (1) on-device edge intelligence, (2)digital twin synchronization, and (3) a generative scenario train-ing layer. Simulated experiments demonstrate that integratingdigital twins and GenAI significantly improves performance –accelerating learning convergence, increasing network through-put, reducing latency, and bolstering resilience against jammingand failures – compared to the original EdgeAgentX. A case studyof a complex tactical scenario (combined jamming attack, agentfailure, and surging network load) highlights how EdgeAgentX-DT maintains operational performance where baseline methodsfalter. The results underscore the promise of digital twin–enabled,generative training for real-world deployment of edge AI incontested environments. The paper is organized into sectionscovering the introduction of the approach, system architecture,methodology, experimental evaluation, discussion of findings, andconclusions. Motivation for Extension:As effective as EdgeAgentXproved, key challenges remain in preparing edge AI for thefull diversity of real-world scenarios. The original frameworkrelieson whatever environments it is exposed to duringtraining – yet the real battlefield can present rare or extremeconditions (severe jamming, sudden node failures, surges intraffic, unforeseen adversarial tactics) that may not be encoun-tered during a limited training phase. Traditional simulators ordomain randomization can provide some variety, but they oftenfall short of covering the vast and evolving space of tacticalscenarios. This is where two emerging technologies becomehighly relevant: digital twins and generative AI.A digital twin is a live, virtual replica of a physical system – in this case, a replica of the communication network andits environment. It mirrors the real network’s topology, state,and behavior in real-time, providing a safe sandbox to analyzeand predict network performance under various situations.Network digital twins have been touted as a revolutionarytool in network management, enabling what-if experimentationwithout risking the actual infrastructure. For tactical networks,a digital twin can simulate battlefield conditions (mobility,terrain, RF propagation) and respond to changes in sync withthe live network. This virtual environment can be used to pre-train and validate edge agents on scenarios that are too riskyor infrequent to test live. Prior research shows that integratinga digital twin into the training loop can markedly improvelearning efficiency and generalization. For example, Duetal.(2023) demonstrate that a DRL agent augmented with adigital twin requires about one-third fewer training interactionsto reach the same performance as a standard DRL agent.The digital twin provides extra, offline training opportunitiesarXiv:2507.21196v1 [cs.LG] 28 Jul 2025 I. INTRODUCTION Modern military communication networks are increasinglyrelying on intelligent autonomous agents operating at the tacti-cal edge. The U.S. Department of Defense has emphasized thatthe tactical edge must be resilient. . .autonomous to executemissions when human oversight is unavailable, and adaptiveto change, underscoring the need for edge AI agents thatlearn and act independently in dynamic, adversarial conditions.EdgeAgentX was recently proposed to address this need viaa novel three-layer framework combining federated learning(FL), multi-agent deep reinforcement learning (MARL), andadversarial defenses. In the original EdgeAgentX, a networkofdistributed edge devices collaboratively learns optimalcommunicati